In the 4S email group, there has been a lot of discussion about Donald Campbell’s two essays “Evolutionary Epistemology” and “Blind Variation and Selective Retention in Creative Thought as in Other Knowledge Processes”. These two essays layout Campbell’s theory about how Popper’s evolutionary epistemology underlies all known knowledge creation processes and explains Campbell’s insights into these ubiquitous knowledge creation processes. Popper himself responded to Campbell’s first essay in “Campbell on the Evolutionary Theory of Knowledge” said approvingly:
Professor Campbell’s remarkable contribution [“Evolutionary Epistemology”] shows the greatest agreement with my epistemology, and (what he cannot know) an astonishing anticipation of some things which I have not yet published when he wrote his paper. (p. 115)“Campbell on the Evolutionary Theory of Knowledge,” by Karl Popper, p. 115 of Evolutionary Epistemology, Rationality, and the Sociology of Knowledge.
All three essays are published in the volume Evolutionary Epistemology, Rationality, and the Sociology of Knowledge. Because these essays aren’t readily available to most, and because they have played such a big role in recent discussions in the 4S email group, I’m going to summarize Campbell’s views here. I unapologetically do this as a reference for future discussion and as such, this is going to be a long post that probably no one will want to read in a single sitting. I plan to do a shorter summary later that will hopefully be more readable.
Where I just give page numbers, the quote is from the essay “Evolutionary Epistemology” by Donald Campbell and the page numbers refer to the book Evolutionary Epistemology, Rationality, and the Sociology of Knowledge. I use “BVSR” is shorthand for the essay “Blind Variation and Selective Retention in Creative Thought as in Other Knowledge Processes” from the same book.
Popper’s Theory of Knowledge: Natural Selection Through Survival of the Fittest Idea
Campbell right away draws the parallel between biological evolution and Popper’s epistemology. He quotes Popper directly as saying:
According to my proposal, what characterizes the empirical method is its manner of exposing to falsification, in every conceivable way, the system to be tested. Its aim is… to select the one which is by comparison the fittest, by exposing them all to the fiercest struggle for survival. … How and why do we accept on theory in preference to others? … We choose the theory which best holds its own in competition with other theories; the one which, by natural selection, proves itself the fittest to survive. … A theory is a tool which we test by applying it, and which we judge as to its fitness by the results of is applications.Karl Popper, p. 49, emphasis mine
Here Popper likens his epistemology to the, at the time, well-known idea of biological evolution being the ‘survival of the fittest’ but in this case the fittest idea rather than the genes fittest at replicating.
Campbell points out that himself Popper went on later to draw parallels between “the trail-and-error learning by man and animals” to Popper’s epistemology. He quotes Popper as saying:
Without waiting, passively, for repetitions to impress or impose regularities upon us, we actively try to impose regularities upon the world. We try to discover similarities in it, and to interpret it in terms of laws invented by us. … These may have to be discarded later, should observation show that they are wrong. This was a theory of trial and error – of conjectures and refutations.Karl Popper, p. 49-50
Popper goes on to explain that since this was true for how we go about discovering regularities via trial and error, that it made sense to him that this should also apply to scientific theories.
I thought that it would apply in the field of science also…. …there is no more rational procedure than the method of trial and error – of conjecture and refutation….Karl Popper, p. 50
Campbell goes on to say “In the process, Popper has effectively rejected the model of passive induction even for animal learning, and advocated that here too the typical process involves broad generalizations from single specific initial experiences.” (p. 50)
All Knowledge Creation Follows the Popper’s Epistemology
Campbell emphasized that all knowledge creation was really identical and all of it followed Popper’s epistemology. Campbell argued that this was true whether we are talking about biological evolution, human ideas, or animal learning.
A focus on the growth of knowledge, on acquisition of knowledge, makes it appropriate to include learning as well as perception as a knowledge process. Such an inclusion makes relevant the learning processes of animals. However primitive these may be, they too must conform to an adequate logical epistemology. Animal learning must not be ruled out as impossible by the logic of knowing.Donald Campbell, p. 52
Campbell quotes Popper for support for this idea:
…I believe, to the growth of pre-scientific knowledge also – that is to say, to the general way in which man, and even animals acquire new factual knowledge about the world. The method of learning by trial and error – of learning from our mistakes – seems to be fundamentally the same whether it is practiced by lower or by higher animals, by chimpanzees or by men of science.Karl Popper, p. 52
Campbell makes it clear what the minimum for evolutionary knowledge creation is — i.e. Karl Popper’s theory:
Three conditions are necessary: a mechanism for introducing variation, a consistent selection process, and a mechanism for preserving and reproducing the selected variation. In what follows we shall look for these three ingredients at a variety of levels.Donald Campbell, BVSR, p. 92
These three elements are simply Popper’s epistemology. Campbell doesn’t even mention “blind variation” in this formulation (yet) because, as we’ll see, that idea simply sheds light on Popper’s existing theory, it is not meant to expand it.
Campbell, in a sense, simply took Popper’s theory and generalized it. Popper was the first to see that the growth of knowledge in science and human theories (explanatory knowledge) used the same process as the growth of knowledge in biological evolution. Campbell noticed that these two examples were not unique and that Popper’s epistemology actually applies to a gigantic hierarchy of knowledge ubiquitous knowledge creation that exists in nature.
A Nested Hierarchy of Selective-Retention Processes
Campbell has now explained, using Popper as his main source, that all knowledge growth comes from simple trial-and-error, that this is the same process as biological natural selection, and that all learning and knowledge – whether human or animal knowledge – use this same process. So far, we’ve not gone any further than Popper’s basic theory generalized to nearly everything.
Human knowledge processes, when examined in continuity with the evolutionary sequence, turn out to involve numerous mechanisms at various levels… hierarchically related, and with some form of selective retention process at each level. (p. 54)Donald Campbell, p. 54
So Campbell’s main thesis is that evolutionary creation of knowledge is ubiquitous and exists at all hierarchies in nature. However, he points out that Popper had this insight before him (though perhaps Popper didn’t apply it quite so broadly.) He quotes Popper as saying:
My theory may be described as an attempt to apply to the whole of evolution what we learned when we analyzed the evolution from animal language to human language. … Neo-Darwinist theory of evolution is assumed; but it is restated by pointing out that its “mutations” may be interpreted as more or less accidental trial-and-error gambits, and “natural selection” as one way of controlling them by error-elimination.Karl Popper, p. 54
His goal is nothing less than to define evolutionary epistemology such as to show that absolutely everything that involves learning and knowledge is really the same process – natural selection via trial-and-error.
Campbell then reveals a new hierarchical concept of nested evolutionary epistemology by quoting Popper on what Campbell calls vicarious selectors.
Our schema allows for the development of error-eliminating controls (warning organs like the eye; feed-back mechanisms); that is controls which can eliminate errors without killing the organism; and it makes it possible, ultimately, for our hypotheses to die in our stead.Karl Popper, p. 54
Note that Campbell is reusing the idea of “ideas dying in our stead” and applying it to the biological and animal world and not just human theories. So this is an interesting new take on this idea. This is key to understanding Campbell’s view. He (and apparently Popper as well) see epistemology as applying to many levels of the organism. Eyes, sense organs, and even classical conditioning are themselves means of utilizing evolutionary processes to eliminate error and create knowledge such that the organism does not have to die to learn.
Campbell notes also that this evolutionary epistemology, according to Popper, applies to even levels of the hierarchy outside of biology.
… Popper [in Clouds and Clocks] has spoken for that emerging position in biology and control theory which sees the natural selection paradigm as the universal nonteleological explanation of teleological achievements… Thus crystals formation is seen as the result of chaotic permutations.. some of which are much more difficult to dislodge than others.
…Turning the model to still lower levels of organization, elements and subatomic particles are seen as but nodes of stability which at certain temperatures transiently select adjacencies among still more elementary stuff.
Turning to higher levels, the model can be applied to such dramatically teleological achievements as embryological growth and wound healing.
Regeneration provides an illustration of the nested hierarchical nature of biological systems. The salamander’s amputated leg regrows to a length optimal for locomotion and survival. The ecological selection system does not operate directly on the leg length however. Instead, the leg length is selected to conform to an internal control built into the developmental system which vicariously represents the ecological selective system. This control was itself selected by the trial and error of whole mutant organisms. If the ecology has recently undergone change, the vicarious selective criterion will correspondingly be in error.
Nested in a hierarchical way within it is the selective system directly operating on leg length, the ‘settings’ or criteria for which are themselves subject to change by natural selection. What are criteria at one level are but ‘trials’ of the criteria of the next higher, more fundamental, more encompassing, less frequently invoked level.Donald Campbell, p. 54-56
Campbell’s Contribution: The Ubiquitousness of Popperian Applications
So if so far all of Campbell’s ideas are also found in Popper, what is Campbell’s contribution? The key insight Campbell adds is the sheer ubiquitousness of how far we can apply Popper’s epistemology. Campbell advocates:
…a systematic extrapolation of this nested hierarchy selective retention paradigm to all knowledge processes, in a way which, although basically compatible with Popper… may go further than he would find reasonable in extremity, dogmatism, and claims for generality.Donald Campbell, p. 56
In other words, Campbell’s theory is a generalization of Popper’s that ubiquitously applies to all novelty at all levels of nature. It goes far deeper than merely looking at biological evolution and human knowledge. To Campbell, all cases of trial-and-error in nature are cases of evolutionary knowledge creation that fall under Popper’s epistemology.
Blind Variation and Selective Retention
As previously quoted, Campbell doesn’t intend to develop a new epistemology. He accepts that Popper’s epistemology is already a correct description of how knowledge is created and that Popper’s epistemology already sets out the minimum requirements for evolutionary knowledge creation.
However, Campbell adds a number of new elements that he believes add insight to Popper’s epistemology, the most important of which he calls “Blind Variation and Selective Retention” (BVSR) which he believes exist in an interplay of hierarchies. Campbell claims:
- A blind-variation-and-selective retention process is fundamental to all inductive achievement, to all genuine increases in knowledge, to all increases in fit of system to environment.
- In such a process there are three essentials: (a) Mechanisms for introducing variation; (b) Consistent selection processes; and (c) mechanisms for preserving and/or propagating the selected variations. Note that in general the preservation and generation mechanisms are inherently at odds, and each must be compromised.
- The many processes which shortcut a more full blind-variation-and-selective retention process are in themselves inductive achievements, containing wisdom about the environment achieved originally by blind variation and selective retention.
- In addition, such shortcut processes contain in their own operation a blind-variation-and-selective-retention process at some level, substituting for overt locomotor exploration of the life-and-death winnowing of organic evolution. (Donald Campbell, p. 56) 
On “Blind” Variation
Campbell here takes an aside to explain one of the most difficult to understand aspects of his theory: why he picked the word ‘blind’ instead of ‘random.’
Campbell recognizes that his understanding of ‘blind’ is probably the most disputed part of his theory: “While most descriptions of discovery and creative processes recognize the need for variation, the present author’s dogmatic insistence on the blindness of such variations seems generally unacceptable.” (Donald Campbell, p. 57)
The key issue here is that Campbell sees his theory as entirely ubiquitous and that wouldn’t be the case if how variations were selected were merely random. In real life, the way trials are chosen and errors are evaluated come to us in many different forms and often they are not random in any meaningful sense. So Campbell has selected the word ‘blind’ to cover as many cases as was possible. “Equiprobability is not needed” he insists. (p. 56)
To be clear how he understands “blind,” Campbell notes that “processes involving sweep scanning are recognized as blind, insofar as variations are produced without prior knowledge of which ones, if any will furnish a selectworthy encounter.” (Donald Campbell, p. 56-57)
So a non-random ‘sweep’ counts as blind variation to Campbell. But he doesn’t stop there. Campbell now attempts to define “blind” more fully and places certain limitations on it:
- “…that the variations emitted be independent of the environmental conditions of the occasion of their occurrence.”
- “… the occurrence of trials individually be uncorrelated with the solution, in that specific correct trails are no more likely to occur at any one point in a series of trials than another, nor than specific incorrect trials.”
- “…rejection of the notion that a variation subsequent to an incorrect trial is a “correction” of the previous trial or makes use of the direction of error of the previous one.” (Donald Campbell, p. 57)
Less Than Blind Variation
And unfortunately, the above 3 limitations seem too limiting for Campbell’s purposes. If taken too literally it would eliminate many parts of the hierarchy Campbell wants to encompass, so he quickly adds an exception that ends up being quite important to understanding his theory:
Insofar as mechanisms do seem to operate in this fashion, there must be operating a substitute process carrying on the blind variations, providing information to the effect that ‘you’re getting warm,’ etc.Donald Campbell, p. 57
Blind Variation = Heuristics that May Fail
So which is it? Are these limitations required for blind variation or not? I feel this is the most difficult part of Campbell’s theory to make sense of. 
I perceive Campbell’s exception as allowing for a variation process that seems less than blind to us (on first glace) so long as the knowledge for how the variations were created was itself created by a blind variation process (one that follows his limitations) at some other level in the knowledge creation hierarchy.
Consistent with my reading of Campbell, Campbell goes on to clarify that the reason seemingly non-blind variations still count as blind variations is because they are based on heuristics (created elsewhere in the hierarchy) that may in fact fail. That is to say, variations can be considered ‘blind’ even if there is neither randomness nor a sweep search so long as their success isn’t guaranteed. They are thus ‘blind variations’ in that they step beyond current knowledge available.
To Campbell, a computer program using heuristics in a deterministic way still counts as “blind” insofar as the heuristics might be wrong. Campbell states this view more explicitly in his discussion on computer algorithms used for knowledge-creation:
For example, one of the heuristics used in Simon’s ‘Logic Theorist’ program is that any substitution or transformation which will increase the ‘similarity’ between a proposition and the desired outcome should be retained as a stem on which further variations are to be tried. Any transformation decreasing similarity should be discarded. …greatly reducing the total search space. It employs an already achieved [at a higher level in the nested hierarchy] partial truth. It produces computer search similar to human problem-solving in failing to discover roundabout solutions requiring initial decrease in similarity.Donald Campbell, p. 67. Emphasis mine.
Here Campbell is using the concept of ‘blindness’ not in how the variations is selected but how they failed to have some selections available that might turn out to be useful later. This is, I believe, the true understanding of Campbell’s use of the term ‘blind variation.’
Campbell goes on to make this view even more clear:
The ‘selectivity,’ [of Simon’s Logic Theorist program] in so far as it is appropriate, represents already achieved wisdom of a more general sort, and as such, selectivity does not in any sense explain an innovative solution. [i.e. it’s ‘blind’] Insofar as the selectivity is inappropriate, it limits areas of search in which a solution might be found, and rules out classes of possible solutions. Insofar as the selectivity represents a partial general truth, some unusual solutions are ruled out. Simon’s ‘heuristics’ are such partial truths…Donald Campbell, p. 66-67. Emphasis mine.
Heuristics created at a different level of the knowledge creation hierarchy that are then used at a lower level of the hierarchy are still considered ‘blind variation’ in Campbell’s theory even if we do not perceive them as such.
Campbell’s Understanding of “Blind Variation” isn’t Intuitive
As a side note, the “code” for Simon’s Logic Theorist is available here. It’s hard to follow, and uses unfamiliar primitives. But as far as I can tell, it used no randomness. Someone correct me if I’m wrong. We’re so used to ‘random’ calls being built in today, but they wouldn’t have existed back then.
This might explain why Simon objected so strongly to Campbell’s concept of ‘blind variation’ and claimed Logic Theorist didn’t use any. A theorem prover that simply tries intelligent things out deterministically and uses no randomness at all strongly breaks our intuition as to what a ‘blind variation’ is. Many people (Simon included) see ‘blind’ as either random or failing to find a good result. Campbell essentially inverses the second of these two intuitive ideas to encompass rules and heuristics that work just fine, but fail to find something interesting outside of its search range. Or, put another way, failing to find even better results than it did due to the inherent limitations in the heuristics being used. This tweak allows Campbell’s epistemology to include deterministic but non-full sweep search algorithms such as Logic Theorist. 
So we’re indebted to this example because it reveals quite a bit about how Campbell understood the phrase ‘blind variation’ differently than most people would upon merely hearing the phrase for the first time. This is also my main criticism of Campbell, that I feel the phrase ‘blind variation’ leaves people with false impressions of what his theory is really saying. There are many algorithms, such as Logic Theorist, that many would claim create knowledge but without ‘blind variation and selective retention’ that Campbell would counter-argue are included under his theory. His disagreement with Simon was a prime example of this. 
Because Campbell sees ‘blind variation’ in such an expansive way, he goes on to explain why he has no concerns trying to reconcile his theory to “an animal in a puzzle box” that is “problem-solving” to get out. He admits that while it appears to us to violate blind variation, this is only because the rules being followed were created at some other level of the knowledge creation hierarchy: “Already achieved wisdom of a general sort which limits the range of trials…” mixed with “vicarious selection” organs (such as vision). (p. 57)
Again, this example strikes me as important in understanding Campbell. An animal in a puzzle box seemingly doesn’t just violate our intuitive sense of blindness but seems to violate all of his criteria for blindness mentioned above. The animal’s trials will not be independent of environmental conditions, its occurrences will not be uncorrelated to the solution, and subsequent variations will be corrections of previous trials. Therefore this example violations his definition of ‘blindness’ on all fronts. This is why I see Campbell’s theory as only requiring these conditions at some level in the hierarchy. When understood in this way — the animal is utilizing knowledge created at some other level of the hierarchy where these conditions did hold — then these criteria make sense. 
This last is how he fits ‘vision’ into his ‘blind variation’ process. At first glance, these are clear opposites. Vision is not blindness by definition. But vision uses knowledge developed elsewhere in the set of evolutionary hierarchies he advocates for. Further, vision isn’t guaranteed to succeed. It’s based on a heuristic that might be stated as something like “you can’t move through things that are opaque to light.” (See p. 60 for his discussion.) We all know this heuristic is often correct, not always. A painful bump into a glass door will quickly demonstrate that vision is itself a ‘blind’ process of sorts.
Blind = Uncertain
So why does Campbell insist on the word “blind” if it is misleading? This has to do with the idea of searching beyond the bounds of existing knowledge:
“…a gain which can only be explained by a continual breakout from the bounds of what was already known, a breakout for which blind variation provides the only mechanism available.” (BVSR, p. 111)
The key aspect of ‘blind’ in ‘blind variation’ is not his randomness, his rules, nor systemic sweep but only that the end result explored into previously unexplored areas. This is why use of heuristics at a higher level in the hierarchy do not violate Campbell’s concept of ‘blind variation.’ 
Identified Hierarchies of Evolutionary Knowledge Creation
Campbell goes on to identify 10 levels (at least) of his hierarchy of evolutionary epistemology:
Non-mnemonic problem solving
Example: a paramecium that uses blind variation of locomotor activity until it finds nourishment. The problem-solutions it finds are then ‘retained’ as a cessation of variation or locomotion. (p. 57-58) Campbell notes that the key element here is (quoting Ashby) ‘…a fundamental tendency for the successful to replace the unsuccessful.” Campbell also points out that even at this level there is a sort of hierarchy. For example, for this system to work there paramecium has to have some sort of ‘sense organ which monitors nutritional level, and substitutes [using Campbell’s concept of ‘vicarious selectors’] for the whole organisms death.
I would note here that this example gives us Campbell’s understanding of what ‘selective retention’ means. Specifically, it doesn’t have to be any more than trying variants until an answer is found and then ‘retaining’ that answer.
Vicarious Locomotor Devices
To address how ‘vision’ is really a case of ‘blind variation’ Campbell starts with a more obvious example: sonar and echolocation. Echolocation is very much like vision but consists of emitting a systematic sweep (which recall Campbell still considers ‘blind’ in that every direction is tried) of sound as a ‘substitute’ to actual locomotion.
He admits that ‘assimilating vision to the blind-variation-and-selective retention model is a more difficult task.” But he suggests that “From the point of view of evolutionary epistemology, vision is just as indirect as radar.” To make this more obvious he suggests imagining a one-photocell substitute eye for the blind where you move it around to ‘see’ light. This process is ‘blind’ in that you are fumbling around trying to find things. Campbell suggests that even a full working eye is really doing the same thing. (p. 59-60)
Campbell points out that vision is a case of vicarious selection that uses knowledge developed at a different level in the hierarchy. As previously mentioned, our vision is based on the contingent discovery that light waves can’t pass through most things that our bodies can’t move through. (See p. 60 for Campbell’s discussion.)
Campbell also points out that we have many different photoreceptors in our eyes and we can think of each of them as different trails. (p. 60.) “Each receptor cell can be conceived of as exploring the possibilities of locomotion in a given direction, the retina collectively thus exploring the possibilities of locomotion in a wide segment of potential directions of locomotion.” (BVSR, p. 95)
“Visual perception seems interpretable as a substitute search process of similar order [to a radar or sonar emitting.]” Though Campbell admits that this analogy is “weakened by the absence of an emitting process on the part of the organism.” (BVSR, p. 95)
Campbell doesn’t specifically mention this, but I’d note that saccades of the eyes are today known to be based on heuristics such as “move the fovea to look at anything that moved because it is likely to be important.”
“These and other considerations convince the present writer that although vision represents the strongest challenge to the generality of a blind-variation-and-selective-retention aspect to all knowledge processes, it is not in fact an exception.” (BVSR, p. 95)
Habit and Instinct: The Necessity of Animal Learning
Here Campbell more directly addresses animal (and some human) learning via habit vs instinct. Despite habit and instinct appearing to be the same thing on the surface, Campbell argues that they are quite different and that the difference is important, though he admits they are “so interlocked and interdependent” that it’s sometimes hard to order them or tease them apart. (p. 60)
But roughly, Campbell defines ‘instinct’ as behaviors wired into the animal via evolution and the genome and ‘habit’ as behaviors learned by the animal and thus is knowledge outside the genome. (i.e. learned via classical conditioning, etc.)
In a crude way, instinct development can be seen as involving a trial and error of whole animals, whereas trial-and-error learning involves the much cheaper wage of responses within the lifetime of a single animal. The same environment is editing habit and instinct development in most cases, the editing process is analogous, and the epistemological status of the knowledge, innate or learned, no different. (p. 60)Donald Campbell, p. 60
But Campbell insists on the importance of this distinction, even if to our eyes they look the same because he points out that it would be impossible for animals to ever develop behavioral instincts via evolution but for the fact that they are also capable of learning habitual knowledge outside their genome:
Complex adaptive instincts typically involve multiple movements and must inevitably involve a multiplicity of mutations at least as greater in number as the obvious movement segments. Furthermore, it is typical that the fragmentary movement segments, or the effects of a single component mutation, would represent no adaptive gain at all apart from the remainder of the total sequence. The joint likelihood of the simultaneous occurrence of the adaptive form of the many mutations involved is so infinitesimal that the blind-mutation-and-selective-retention model seems inadequate.Donald Campbell, p. 61
But Campbell points out that if trial and error learning precedes instinct, then this poses no problem at all:
The adaptive pattern being thus piloted by learning, any mutations that accelerated the learning, made it more certain to occur, or predisposed the animal to certain components responses, would be adaptive and selected no matter which component, or in what order affected. The habit thus provided a selective template around which the instinctive components could be assembled. (Stating it in other terms, learned habits make new ecological niche available, which niche then selects instinct components.) It is furthermore typical of such instincts that they involve learned components, as of nest and raw materials location, etc. (p. 61)Donald Campbell, p. 61
Visually Supported Thought
By this Campbell means specifically use of eyes (as previously discussed) as a “substitute trail and error of potential locomotion…” (p. 62)
Mnemonically Supported Thought
Here Campbell really means using one’s memory in place of the outside environment to try out “thought trails being selected by vicarious criterion substituting for an external state of affairs. The net result is the ‘intelligent’, ‘creative’, and ‘foresightful’ production of thought…” (p. 62)
Knowledge Creation Through Computer Algorithms
Campbell here makes an aside to claim that computer problem solving (the term “AI” didn’t exist yet) is also often a case Mnemonically Supported Thought. He admits that this isn’t obvious and that it’s a contended position, but makes the bold claim:
“The present writer would insist here too, that if discovery or expansions of knowledge are achieved, blind variation is requisite.” (p. 66)
This part of Campbell’s essay puts him squarely at odds with David Deutsch’s assertion that artificial evolution does not produce knowledge. (See BoI p. 160-161) Indeed, Campbell is going the opposite direction: trying to claim that computer programs that create knowledge are common because they are secretly utilizing his blind variation and selective retention process (though this might not be obvious as first as discussed above.)
Campbell, elsewhere, includes a computer program playing chess as an example of a computer program creating knowledge using a Popperian knowledge creation process. Such a program, tries out blind variations and builds heuristics at various levels as to where it should try more blind variations if the potential move looks promising. To Campbell, this is a clear case of blind-variation-and-selective-retention at level of a hierarchy. (BSVR, p. 106)
Campbell addresses criticisms of his view. He quotes Herbert Simon (a contemporary computer scientist) as “rejecting his point of view, at least in the extreme form advocated here.” (p. 66)
The more difficult and novel the problem, the greater is likely to be the amount of trail and error required to find a solution. At the same time, the trail and error is not completely random or blind; it is, in fact, highly selective.Hebert Simon, p. 66
But Campbell defends his view against Simon by arguing (as I previously quoted in part) that:
The ‘selectivity,’ in so far as it is appropriate, represents already achieved wisdom of a more general sort, and as such, selectivity does not in any sense explain an innovative solution. Insofar as the selectivity is inappropriate, it limits areas of search in which a solution might be found, and rules out classes of possible solutions. Insofar as the selectivity represents a partial general truth, some unusual solutions are ruled out. Simon’s ‘heuristics’ are such partial truths, and a computer which would generate its own heuristics would have to do so by a blind trial and error of heuristics principles, selection from which would present achieved general knowledge.Donald Campbell, p. 66-67
If the computer program generates its own learned heuristics in its search for an innovative solution Campbell does not see this as anything but “similar to human problem-solving in failing to discover round-about solutions requiring initial decreases in similarity. Beyond thus applying what is already known, albeit only a partial truth, the new discoveries must be produced by a blind generation of alternatives.” (p. 66-67)
Socially Vicarious Exploration: Observational Learning and Imitation
This is essentially social learning. “In this, the trial-and-error exploration of one member of a group substitutes for, renders unnecessary, trial-and-error exploration on part of other members. The use of trial and error by scouts on the part of migrating social insects and human bands illustrates this general knowledge process. At the simplest level in social animals are procedures whereby one animal can profit from observing the consequences to another of that other’s acts….” (p. 67)
By “language” Campbell includes animal “language” such as bees. A bee communicates via its dance knowledge that is “tremendously reductive of flight lengths on the part of the observing or listening worker bees.” (p. 68)
He points out that ant and termite pheromones play a similar role of ‘language’ he intends.
…an explorer who has encountered food exudes a special external hormone… The other workers backtrack on this special scent. If they too are successful, if the food supply remains successful, they keep the pheromone track renewed. (p. 69)Donald Campbell, p. 69
In other words, knowledge is communicated via “language” as a vicarious selector in place of having to learn the same knowledge via trail-and-error yet it is still a case of blind variation (the food may or may not actually be there) and selective retention (the path diminishes over time if not renewed) in that the knowledge being communicated will change over time as the situation changes.
Campbell also extends all this to how humans use language. When we communicate via language the whole process is a blind variation / selective retention process by which each word or sentence has multiple possible meanings and we have to eliminate possible interpretations via the context of the whole series of the communication. (p. 69)
To a Popperian audience, I don’t need to explain this one. Though Campbell didn’t have the word “memes” to work with, he speaks of the ways that social organizations compete and how we adapt other people’s idea via selective borrowing. (p. 70)
And finally Campbell reiterates the blind-variation-and-selective-retention in the scientific method.
To Campbell, knowledge creation was ubiquitous and exists in interacting hierarchies where knowledge from one hierarchy can be used in the hierarchies below. Campbell defined this process as a process of “blind variation and selective retention” though he intended this to be identical to Popper’s epistemology of knowledge creation through trail-and-error or conjecture-and-refutation.
Campbell’s contributions to Evolutionary Epistemology were heavily rooted in Popper’s already existing theory. In many ways, Campbell is just restating what Popper had already worked out. However, Campbell insisted that even Popper wasn’t catching just how ubiquitous was knowledge-creation. Whereas Popper was primarily interested in explanatory knowledge and human theories, Campbell conjectured that there is a vast hierarchy in nature of knowledge creation and that every level of that hierarchy utilized Popper’s epistemology of trail-and-error and retaining solutions that worked.
Campbell also made the contribution that through organs that are “vicarious selectors” biological evolution discovered ways to allow our senses and our minds to try out actions vicariously that allowed organisms to let their ‘ideas die in their place.’
 On Science and “Induction“
Lest we think Campbell has slipped into the error of induction, his footnote says the following:
The use of the phrase “inductive achievement” is for convenience in communicating and does not in the least imply advocacy of the Bacon-Hume-Mill explanation of those achievements nor disagreement with Popper’s brilliant criticisms of induction.”(Donald Campbell, p. 56)
I would note here that this is the true Popperian view of the use of the term “induction.” Popper did not believe in Word Essentialism and so he had no problem with referring to science as ‘inductive achievement.’ Not only does Popper does not object to Campbell’s use of the term “induction” as meaning simply knowledge creation via science, but Popper himself states that the word “induction” may well be repurposed to refer to Popper’s own theory:
I never quarrel about words, and I have of course no serious objection if you wish to call the method of critical discussion ‘induction.’Karl Popper, Myth of the Framework, p. 104
Popper simply asks that we recognize how that new use of the word differs from previous uses. Popper concentrated on concepts rather than words.
 On Blind Variation
I acknowledge that this is the area Ella, in particular, is going to disagree with my reading of Campbell. This post just attempts to collect the necessary references and quotes that explain my understanding of Campbell and not argue this point further yet. I will save that for future posts.
However, I hope this post, particularly the Logic Theorist example, explains why I feel she is misunderstanding Campbell’s intent and why I feel little doubt that Campbell would instantly see Gradient Descent (GD) as a form of ‘blind variation.’ In my view, Campbell would see Gradient Descent’s outer loop (the hyperparameter search) as one level of the hierarchy of the ‘blind variation’ process that creates the heuristics than used by the GD algorithm to generate its variations. Since the heuristics can fail (and in fact always do at first until you find the right hyperparameters) GD itself also counts as a blind variation process for the very same reasons that Campbell argued that Logic Theorist does. The paramecium example also explains why I have little doubt he’d see Gradient Descent as a form of ‘Selective Retention.’
I hope this clarifies why I’m struggling to see her reading (which declares GD not a BVSR process) as compatible with the actual examples that Campbell uses in his papers. However, I would encourage Ella to do her own post here to clarifying her own reading of Campbell if it is contrary to mine. It’s always possible I’m missing something important. Also, I admit, that Campbell is not an ‘authority’ and it may well be that I’m reading Campbell correct but Campbell’s theory is wrong. (A possibility I consider in next week’s post where I look at Dennis’ view of the subject.) So I hope Ella will continue her conversations with me using this post as the ‘next step’ in the conversation.
 Logic Theorist and Randomness
Here I’m acting as if I know that Logic Theorist used no randomness, when in fact I am not sure. But in any case, even if Logic Theorist did use some randomness, Campbell was apparently not aware of this, or that would have been an easier argument to make. So my argument is based on Campbell’s (perhaps incorrect) understanding of Logic Theorist as a non-random program, thus forcing him to invert ‘blindness’ from failing to find a good result to failing to find an even better result.
 Animal in a Puzzle Box Examples
To be more clear here, an animal in a puzzle box might try something and fail to get good leverage, so its next trial corrects that. Or it might look around the environment and use clues from the environment on what to try next. Or it might, through classical conditioning, learn that certain things give rewards and others don’t so it’s trials start to correlate with the solution. In each of these cases, the animal violates Campbell’s criteria for blindness in its trials. But Campbell would argue that none of this matters because the knowledge created to allow for these ‘violations’ where themselves created at a different level of the hierarchy and at that level the rules held.
So, for example, the reason the animals can use clues in the environment is due to its vision, which is itself just a bunch of photocells in the eye — each of which is a blind trial uncorrelated to the environment. And when an animal corrects its leverage, this is due to biological knowledge that came from biological evolution. Likewise, classical conditioning came from biological evolution. Since biological evolution follows his rules, then he’d claim the whole idea of an animal escaping from a puzzle box doesn’t violate his rules as a whole.